As of 13 August 2026, AI can only partly clean your customer data.

What to actually do

1. Use a tool built for this

   The route this page recommends

2. Do it yourself

   Second choice

   A chat interface, chat-fluent skill, and roughly 1 hour until you can act on the result.

   How to actually do it

   1. Open the source spreadsheet or CRM export and make a read-only backup with the original row identifiers and column names intact.
   2. Write down the meaning of each column, the authoritative customer identifier, the permitted formatting changes and the exact rule for treating two records as duplicates.
   3. Confirm that your organisation permits the chosen AI service to process this data; if it does not, create an anonymised sample or use an approved workplace tool instead.
   4. Paste the prompt into the AI tool, attach the approved file or anonymised sample, and supply the column definitions and cleaning rules in the bracketed sections.
   5. Export the proposed-change list and compare its row identifiers, original values and new values against the source file; reject any guessed value, unsupported merge or forbidden change.
   6. Ask the AI to produce the approved formulas, SQL or transformation script, then run it on a copy rather than the original data.
   7. Compare before-and-after row counts, unique identifiers, missing required fields, duplicate counts and linked-system totals, and have the data owner inspect ambiguous matches.
   8. Apply only the approved changes to the live dataset, retain the original backup and record which rules and exceptions were used.

   Prompt

   I need to clean a customer-data file for internal business use.
   
   Use the attached file or the sample below only if it is approved for this AI service. Do not expose, repeat or infer personal information unnecessarily. If the file contains identifiable customer data and this service is not approved for it, stop and tell me to use an anonymised sample or an approved workplace tool.
   
   Data context:
   - Source system: [CRM, spreadsheet or other source]
   - Intended destination: [destination system]
   - Meaning of each column: [paste the column names and definitions]
   - Authoritative fields: [for example, customer ID or confirmed email]
   - Allowed changes: [for example, trim whitespace, standardise phone formats, fix capitalisation, flag duplicates]
   - Forbidden changes: [for example, changing customer IDs, guessing missing values, deleting rows]
   - Duplicate rule: [state exactly when two records may be treated as the same]
   - Required formats: [dates, postcodes, telephone numbers, email addresses and other rules]
   
   First, profile the data and report row count, column count, missing values, duplicate candidates, inconsistent formats, invalid values and values outside the stated rules. Do not silently change anything.
   
   Then produce:
   1. A cleaning plan explaining each proposed change.
   2. A list of every proposed change with row identifier, column, original value, new value and reason.
   3. A separate list of ambiguous duplicate matches and records needing human decisions.
   4. A reproducible spreadsheet formula, SQL statement or transformation script for each mechanical change.
   5. Validation checks for row counts, unique identifiers, required fields, referential links and before-and-after totals.
   6. A cleaned output only after I explicitly approve the proposed changes.
   
   Never invent missing customer details, merge records solely because names look similar, delete records to make totals match, or claim that an invalid value is correct without evidence.

   Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.

3. Hand it to a person

   The distant third

   A person who owns the outcome does this end to end, worth it when the failure is dear.